1/I built an open-source prompt injection detector. DeBERTa-v3-base trained on a curated dataset of 54K samples.
96.1% detection rate at 0.1% false positive rate — same architecture as ProtectAI v2, but +17pp from better training data.
4/It's not a silver bullet. Jailbreak detection is the weakest at 93.5%. English-dominant. Single-turn only. And at real-world attack rates (<0.1%), even 0.1% FPR means false alarms outnumber real detections.
It's Layer 1 in a defense stack — fast pre-filter, not the final word.
5/Everything is open source:
Dataset: huggingface.co/datasets/dmil… Model: huggingface.co/dmilush/shiel… Code: github.com/dvm81/shieldlm Writeup: medium.com/@dimiter.milushev…
Dimiter Milushev retweeted
NeRD and Neural-PIL enable decompositions into shape, BRDF, and lighting. However, they require camera poses. Common in-the-wild collections are not only in multiple illuminations but also locations. Here, COLMAP fails. SAMURAI aims to fix this: markboss.me/publication/2022…
🧵
Dimiter Milushev retweeted
My non-inclusive tweet for the year: I'm looking forward to the time when all of the new year's resolution gym members realize a certain amount of commitment is needed and most stop going. It's far too crowded during the first few weeks of a new year.
Dimiter Milushev retweeted
Excited to share our #emnlp2018 work: SOTA on many NLP tasks where performances already feel saturated! ideas: models self-train on unlabeled data by refraining access from full representations of text sequences + multitask-traing. by @clark_kev, @chrmanning, @quocleix & I.
Cross-View Training—A semi-supervised learning technique by @clark_kev @lmthang Quoc Le @chrmanning at #emnlp2018. Allows training for your #NLProc task on large-scale unannotated data, not only using such data for task-agnostic word representations arxiv.org/abs/1809.08370
Really cool xitter.cf/seb_ruder/status/10062…
Transfer learning with language models is getting hot! 🔥New state-of-the-art results today by two different research groups: Trinh and Le (Google) on the Winograd challenge and Radford et al. (OpenAI) on a diverse range of tasks.
arxiv.org/abs/1806.02847
blog.openai.com/language-uns…
Dimiter Milushev retweeted
"I interviewed at 5 top companies in 5 days, and got 5 job offers" - honest article about interviews & preparation medium.com/@XiaohanZeng/i-in…
Really nice talk! xitter.cf/sansforensics/status/8…
Toppling the Stack: Outlier Detection by @davidjbianco #threathunting #threathuntingsummit 2017 TALK NOW... youtu.be/7q7GGg-Ws9s?a
Dimiter Milushev retweeted
My plane was delayed, so this is the setup for my #fsharpconf talk in 20 minutes... #hoorayforairportwifi
Dimiter Milushev retweeted
Джон Гришам пуска безплатно новата си книга, защото смята, че е най-важната, написана от него. "Туморът",... fb.me/3M9E0PSQk
Dimiter Milushev retweeted
Черен ден в българската история, помните! А защо не е в учебните програми - отговорете си сами! fb.me/28COhOlXD
Dimiter Milushev retweeted
@France_Bulgarie: "Има хора, които не могат да разберат, че има други хора, които могат да работят само за заплатата си, за честта"